Skip to content
HumanizerOps CONTROL PLANE
$428 / $700

RUNS / JUL-14 / RUN-7F31A9

Transformation evidence report

partial evidence

Technical article · 3 providers · 4 detectors · policy production-safe@24

RUN ID
RUN-7F31A9
SOURCE
sha256:b7e9…23af
POLICY / RECIPE
v24 / editorial-safe@7
EXECUTED
14 Jul 2026 · 18:42:09
COST
$0.1842 / $0.25 max
LATENCY
12.84s / 20s max
RETENTION
◆ HASH ONLY

SAFE SELECTED · EDITORIAL REVIEW RECOMMENDED

Atlas Edit / atlas-3.2-long

Candidate C-017 passes every hard constraint. One detector timed out, so the result is selected on available evidence—not declared best overall.

SEMANTIC
0.947
PASS ≥ 0.88
DETECTOR SIGNAL
0.182
SECONDARY
HARD GATES
8/8
ALL PASSED
REPORT
STATIC P0
REDACTED

Missing evidence: Detector Northstar-1 timed out at 8.0s. No score was imputed.

1safe selectedof 3 candidates
8/8hard constraintssource protected
1detector reviewsignal is secondary
1missing evidencetimeout retained
14mfreshnessprovider registry

GUIDED INTAKE

Replay benchmark

✓ READY
SOURCE
Saved artifact set
PRIVACY
Hash only
REPORT
Redacted
HARD STOP
Number or entity loss

PROVIDER ELIGIBILITY

3 eligible 1 missing connector
No document text is written to analytics or shared logs.

PROVIDER MATRIX

Candidate decisions

3 candidates / 3 providers
Provider attempts and candidate evidence
Decision Provider / model Semantic Detector Protected Cost Latency Reason Freshness
✓ safe selected Atlas Editatlas-3.2-long · eu-west 0.947pass 3/41 timeout 8/8 passclaims + entities $0.071 8.42s All hard gates passed 14m
≈ best observed Local Editorialmistral-small-3.1 · sto-1 0.919pass 4/4lowest signal 8/8 passclaims + entities $0.042 4.18s Voice fit below target 14m
× rejected Vector Prosevp-rewrite-2 · us-east 0.861hard fail 2/42 unsupported 7/8 failnumber removed $0.071 12.84s Number mismatch 22m

Humaniser engine / live

Analyze, revise, verify

Source passage

0 words

Profiles are explicit, bounded transformations. No profile guarantees a detector outcome.

Evidence

Waiting for input

Run a live analysis to see model-specific evidence and revision options.

Detector output is probabilistic evidence, not proof of authorship. Source text is excluded from KlarAds and PostHog events.